Model

Hunyuan-Large (A52B)

OPEN

Tencent · Hunyuan family · released Nov 5, 2024

389B-total/52B-active open-weights MoE, the largest open-source Transformer-based MoE at release; beats Llama-3.1-405B on MMLU, GSM8K, MATH and HumanEval.

ReasoningCodingVisionFunction callingTool useAgentic

1072.5

Elo · rank #197

Parameters

389B

Active params

52B (MoE)

Context

128K tokens

Architecture

Sparse Transformer-based MoE (389B total / 52B active)

License

Tencent Hunyuan Community License Agreement

Languages

—

API price (in/out)

No hosted API

Modalities

text

Benchmark results

Bar shows position within the tracked field; marker = field best

AlpacaEvalHuman preference51.8%#15

best: Tülu 2+DPO 70B · 95.1%

ARC-ChallengeReasoning95.0%#6

best: Llama 3.1 405B · 96.9%

Arena-HardHuman preference81.8%#24

best: Qwen3-235B-A22B (Non-Thinking) · 96.1%

BIG-Bench HardReasoning86.3%#13

best: ERNIE 4.5 300B-A47B · 94.3%

C-EvalKnowledge91.9%#6

best: Qwen3.6-Plus · 93.3%

CMMLUKnowledge90.2%#3

best: Doubao-1.5-Pro · 90.9%

DROPReasoning88.9#9

best: Hunyuan-T1 · 93.1

GPQA DiamondReasoning42.4%#208

best: GPT-5.6 · 94.6%

GSM8KMath92.8%#25

best: Llama 3.1 405B · 96.8%

HellaSwagReasoning86.8%#21

best: Claude 3 Opus · 95.4%

HumanEvalCoding71.4%#80

best: Claude Opus 4.5 · 99.4%

IFEvalReasoning85.0%#59

best: Gemma 4 26B A4B · 98.5%

MATH-500Math69.8%#92

best: GPT-5 · 99.4%

MBPPCoding72.6%#31

best: Llama-3.3-Nemotron-Super-49B v1 (Reasoning On) · 91.3%

MMLU-ProKnowledge60.2%#115

best: Claude Fable 5 · 91.5%

MMLUKnowledge88.4%#25

best: OpenAI o3 · 92.9%

MT-BenchHuman preference9.4#1

best: this model · 9.4

PIQAReasoning88.3%#4

best: GPT-4o mini · 93.1%

TriviaQAKnowledge89.2%#3

best: Sarvam-1 (2B) · 90.6%

WinoGrandeReasoning88.7%#4

best: PaLM 2 · 90.9%

Run it locally

VRAM @ Q4

215 GB

VRAM @ FP16

780 GB

Fits on (Q4)

M3 Ultra 512GB

389B total params vastly exceed a single RTX 4090's 24GB VRAM even at 4-bit; requires multi-GPU or heavy CPU offload to serve.

Quantizations

GGUF Q4

Hunyuan family

Elo progression across releases

Hunyuan (1st-gen)Sep 20231000.0 Hunyuan-Large (A52B)Nov 20241072.5 Hunyuan-TurboSFeb 20251306.0 Hunyuan-T1Mar 20251631.0 Hunyuan-A13BJun 20251517.1 Hy3Jul 20262485.4

Resources

HFtencent/Tencent-Hunyuan-Large GHTencent-Hunyuan/Tencent-Hunyuan-Large APIAvailable via Tencent API

API price weights · each benchmark row carries its own source badge (see methodology)